Add NemotronH to KNOWN_BASES; handle llm_config nesting pattern

scan.py

@@ -66,6 +66,11 @@ KNOWN_BASES = {
         "vocab_size": [256000, 262144],
         "model_type_patterns": ["gemma"],
     },
+    "nemotron_h": {
+        "name": "NemotronH (NVIDIA Mamba+MoE hybrid)",
+        "vocab_size": 131072,
+        "model_type_patterns": ["nemotron_h", "nemotronh"],
+    },
 }
 
 
@@ -137,11 +142,16 @@ def stage1_screen(model_id: str, config: dict) -> dict:
     Returns a structured verdict without downloading any weights.
     Handles nested text_config (Qwen3.5/3.6, Mistral3, MiMo-V2.5 pattern).
     """
-    # Merge text_config into top-level if present (multimodal nested configs)
-    if config.get("text_config") and not config.get("vocab_size"):
-        tc = config["text_config"]
-        config = {**tc, **{k: v for k, v in config.items()
-                           if k not in ("text_config", "vision_config", "audio_config")}}
+    # Lift nested LLM config into top-level when top-level vocab/hidden is absent.
+    # Handles: text_config (Qwen3.5/3.6, Mistral3, MiMo-V2.5), llm_config (NemotronH Omni)
+    _NESTED_KEYS = ("text_config", "llm_config")
+    _SKIP_KEYS = ("text_config", "llm_config", "vision_config", "audio_config", "sound_config")
+    if not config.get("vocab_size"):
+        for nested_key in _NESTED_KEYS:
+            if config.get(nested_key) and config[nested_key].get("vocab_size"):
+                tc = config[nested_key]
+                config = {**tc, **{k: v for k, v in config.items() if k not in _SKIP_KEYS}}
+                break
 
     vocab = config.get("vocab_size")
     model_type = (config.get("model_type") or "").lower()